Raphexon's NNUE eval net

[kranium]

Raphexon (Gekkehenker)
has been training NNUE nets for many weeks

He's very active in this regard on sf-nnue discord
and has created several very strong ones...

Anyone looking for an alternative to StockfiNN to try/test
can download his latest (GK2020-07-12.7z) here:

https://discord.com/channels/4359437104 ... 3457660929

Note: I have not yet tested this.

Thanks Henk!

[matejst]

Thanks, Norman. I managed to register on Discord and found an invitation to the channel. The net seems very good indeed, and new binaries are faster each day.

[peter]

Raphexon (Gekkehenker)
has been training NNUE nets for many weeks

He's very active in this regard on sf-nnue discord
and has created several very strong ones...

Anyone looking for an alternative to StockfiNN to try/test
can download his latest (GK2020-07-12.7z) here:

https://discord.com/channels/4359437104 ... 3457660929

Note: I have not yet tested this.

Thanks Henk!

Thanks Norman and thanks Henk!

Of course some faster than the bigger ones with same binaries as for n/s, which doesn't mean even as much as it does for SF dev., time to solution is the only thing mattering here, time to depth doesn's even count as much as for SF dev. neither, as it seems to me.

One question I apologize already in advance as for the lack of basic knowledge about the matter. Couldn't a binary of it's own make an existing net being trained at gameplaying and at analysing of .pgns at once? Why is a training procedure of it's own necessary at all, if a net has reached a certain amount of data already, playing head to head with SF with its hard coded eval? Wouldn't be training the net by gameplaying against other engines helpful too?

[matejst]

Peter,

I think that Pleomati, on Discord, has implemented a learning function for StockFish NNUE and had good results with it. I do not know how this learning functions, if it is a way of training the NN or something more classical. Indeed, an implementation of persisting learning/training would be quite interesting.

[Raphexon]

Non-default slowmover no longer seems needed.
I no longer have time issues with either Pleomati's binaries or Nodchip's latest from 7-13.

In fact, my 27-06 net is fully competitive with a month old SF-dev* now, even at 10s+0.1s.


*Haven't tested against the very latest, but should be close anyway.

[peter]

Peter,

I think that Pleomati, on Discord, has implemented a learning function for StockFish NNUE and had good results with it. I do not know how this learning functions, if it is a way of training the NN or something more classical. Indeed, an implementation of persisting learning/training would be quite interesting.

Thanks!
Will give this one
https://discord.com/channels/4359437104 ... 9589460992
a try.

[peter]

Good fast binaries, but no learning at playing.

[matejst]

Peter,

You have to search in one of the threads, where Pleomati mentioned that SF NNUE did not lost a game in the last 70. I think he did not upload these binaries, but a bit modified NodChip binaries, and indeed extremely fast binaries.

[peter]

You have to search in one of the threads, where Pleomati mentioned that SF NNUE did not lost a game in the last 70. I think he did not upload these binaries, but a bit modified NodChip binaries, and indeed extremely fast binaries.

Thanks again, I'll try to talk to Pleomati at Discord.

[matejst]

Peter, try here: https://discord.com/channels/4359437104 ... 4093398056